Current search methods have a weakness in resampling particles at fixed temperature that ignores reward distribution across denoising steps

the search then resamples particles at a fixed temperature that ignores how rewards spread across each denoising step

http://arxiv.org/abs/2608.26585v1